From indicators to biology: the calibration problem in artificial consciousness

A new working paper by researcher Florentin Koch, titled 'From indicators to biology: the calibration problem in artificial consciousness,' delivers a sharp critique of current efforts to evaluate consciousness in AI systems like GPT-4 or Claude 3.5. Koch argues that while the field has progressed beyond simplistic behavioral tests (like the Turing test) to analyze internal architecture and use theory-derived indicators, this approach remains 'epistemically under-calibrated.' The core problem is a lack of scientific consensus on consciousness itself, the absence of independent validation for proposed indicators, and—critically—no 'ground truth' of artificial phenomenality against which to measure. Under these conditions, Koch asserts that any probabilistic attribution of consciousness to current large language models or other AI is premature and scientifically unsound.

Instead of continuing down this uncertain path, Koch proposes a strategic pivot for near-term research. He advocates redirecting effort toward biologically grounded engineering approaches that narrow the gap with living systems, the only domain where consciousness is empirically observed. This includes developing biohybrid systems (combining biological and artificial components), neuromorphic computing (hardware that mimics neural structures), and connectome-scale simulations (modeling brain wiring at a massive scale). The paper suggests that by tethering AI development more closely to the known biology of natural consciousness, the field can build a more defensible foundation for future inquiry, moving from speculative indicators toward engineered systems with clearer biological plausibility.